[MaartenMarsman]

Handoff finding F5 (dev/audit/2026-10-16-prior-sensitivity-review-handoff.md).

wobble_yardstick() selected threshold-relevant edges (|log BF| <= 3 log 10) on the first refit alone. An edge that saturated in the replicate but not in the first refit could carry the noise band up, and the reverse case was dropped without being counted. The band now pools the edges inside the window in both refits, and an edge inside it in only one of them, saturated or finite beyond the window, is counted as censored whichever refit that is. The tutorial's Appendix D already describes this rule.

Changes: R/prior_sensitivity.R (the function, its roxygen, the printed censored line), the regenerated Rd, one sentence in vignettes/prior-sensitivity.Rmd, two tests for the swapped-input symmetry, and a NEWS entry saying the band and the within-noise classification can change for edges near the thresholds.

Reproduction (dev/review-2026-08/repro/run-f5.sh, develop vs this branch):

• the handoff's example, log(c(2, 2)) vs log(c(2000, 3)): develop gives a band of 6.58 one way and 0.41 the other, censored 0 both ways; this branch gives 0.41 both ways, censored 1.
• an edge at |log BF| 2.30 in the first refit and 9.21 in the replicate: develop pools it (band 5.94); this branch censors it (band 0.46, censored 1).
• a small Wenchuan check whose only censored edge saturates in the replicate: identical band, count and mover classification on both sides; only the printed wording of the censored line differs.

test-prior-sensitivity.R on the branch: 104 pass, 0 fail (every-run tier).